Accelerating Open-Source InSAR Time-Series Analysis on Consumer GPUs: A PyTorch-Based Network Inversion Solver for MintPy
Year: 2026
Type: article
Abstract: Poster presented at TASTI 2026 (Taiwan International Assembly of Space Science, Technology, and Industry), Tainan, Taiwan; poster session on 10 November 2026, Paper No. T0043. InSAR time-series analysis measures slow ground deformation, and the SAR archives it consumes keep growing. In a complete run of MintPy, a widely used open-source small-baseline package, a single step, the network inversion, takes 43.5% of the total step time (Fernandina scene, CPU). It is a weighted least-squares problem solved pixel by pixel. Porting it to the GPU as it is, with PyTorch's QR-based least-squares solver, gave only 1.43x over the CPU path: at this matrix height the solver works through the pixels one at a time. Recasting every pixel's system as normal equations turns the work into batched matrix products and a batched Cholesky factorization in single precision, written in PyTorch without custom GPU code. On one consumer GPU (NVIDIA GeForce RTX 5080, 16 GB), the inversion of a Galápagos scene (3.4 million pixels, 475 retained interferograms) fell from 6,169 s to 139 s, 44.4x as timed by MintPy with interpreter start-up excluded, within 7.6 GB of GPU memory. Squaring the condition number in single precision leaves micrometre-level differences from the CPU's SVD-based solver, buried under InSAR's other error sources, and all five public test datasets pass a numerical agreement gate against the CPU path. The solver is submitted to MintPy as pull request #1490 (under review). Poster sources, figures and benchmark notes: https://github.com/s-sasaki-earthsea-wizard/tasti2026-poster
Author: Sasaki, Syota
Language: English
Subfield: Aerospace Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7223669048